bpr.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# -*- coding: utf-8 -*-\n",
    "\n",
    "import torch\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "\n",
    "dtype = torch.float\n",
    "device = torch.device('cpu')\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "rating = np.load('the-movies-dataset/numpy/small_ratings.npy')\n",
    "users = np.load('the-movies-dataset/numpy/users.npy')[:100]\n",
    "movies = np.load('the-movies-dataset/numpy/movies.npy')[:100]\n",
    "\n",
    "users_num, movies_num, k = len(users), len(movies), 5\n",
    "rating_len = len(rating)\n",
    "\n",
    "print(rating[:, 2].max())\n",
    "print(rating[:, 0].max())\n",
    "print(len(users)-1)\n",
    "print(rating_len)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# normalization\n",
    "rating[:, 2] -= 2.5\n",
    "rating[:, 2] /= 2.5\n",
    "\n",
    "# thita\n",
    "W = torch.randn(users_num, k, device=device, dtype=dtype) / 10\n",
    "H = torch.randn(movies_num, k, device=device, dtype=dtype) / 10\n",
    "\n",
    "print(rating[:, 2].max())"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Create Ds"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ds = []\n",
    "for ui, i, ri in rating:\n",
    "    for uj, j, rj in rating:\n",
    "        if ui != uj or i == j:\n",
    "            continue\n",
    "        if ri > rj:\n",
    "            ds.append((int(ui), int(i), int(j)))\n",
    "\n",
    "ds = list(set(ds))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "ds[np.random.randint(len(ds))]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict(u, i):\n",
    "    Wu = W[u].view(1, W[u].size()[0])\n",
    "    return torch.mv(Wu, H[i])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def predict_diff(u, i, j):\n",
    "    return  (predict(u, i) - predict(u, j))[0]\n",
    "\n",
    "print(predict_diff(0, 0, 1))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def partial_BPR(x_uij, partial_x):\n",
    "    exp_x = np.exp(-x_uij)\n",
    "    return exp_x / (1 + exp_x) * partial_x"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "iteration = 100000\n",
    "lr = 1e-4\n",
    "\n",
    "def train(W, H, lr=1e-3, rr=0.02):\n",
    "    for itr in range(iteration):\n",
    "        u, i, j = ds[np.random.randint(len(ds))]\n",
    "        x_uij = predict_diff(u, i, j)\n",
    "\n",
    "        for f in range(k):\n",
    "            W[u][f] -= lr * (partial_BPR(x_uij, H[i][f] - H[j][f]) + rr * W[u][f])\n",
    "            H[i][f] -= lr * (partial_BPR(x_uij, W[u][f]) + rr * H[i][f])\n",
    "            H[j][f] -= lr * (partial_BPR(x_uij, -W[u][f]) + rr * H[f][f])\n",
    "        \n",
    "        if itr % 10000 == 0:\n",
    "            print(W[18])\n",
    "            print(H[0])\n",
    "            \n",
    "    return W, H\n",
    "\n",
    "W, H = train(W, H, lr)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(W[18])\n",
    "print(H[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(W[18])\n",
    "print(H[0])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Evaluation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "user_id = 18"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "Wu = W[user_id].view(1, W[user_id].size()[0])\n",
    "prediction = list(zip(list(range(movies_num)), torch.mm(Wu, H.t()).tolist()[0]))\n",
    "prediction.sort(key=lambda x: x[1], reverse=True)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### get user preference"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "movie_rates = []\n",
    "movie_predict_rates = []\n",
    "\n",
    "for u, i, r in rating:\n",
    "    if u == user_id:\n",
    "        movie_rates.append((int(i), r))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### read movies' titles and genres"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import json\n",
    "movie_data = []\n",
    "df = pd.read_csv('the-movies-dataset/movies_metadata.csv')\n",
    "\n",
    "for index, row in df.iloc[:, [3, 8]].iterrows():\n",
    "    movie_data += [{'title': row['original_title'], 'genres': [x['name'] for x in json.loads(row['genres'].replace('\\'', '\"'))]}]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### print result"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print('User ', users[user_id])\n",
    "print('from rating, he/she likes:')\n",
    "print('%s %25s %43s' % ('movie_id', 'movie_title', 'movie_genres'))\n",
    "for m, r in movie_rates:\n",
    "    if r > 0.5:\n",
    "        mid = movies[m]-1\n",
    "        print('%8s %25s %43s' % (mid, movie_data[mid]['title'][:24], movie_data[mid]['genres'][:4]))\n",
    "\n",
    "print('')\n",
    "print('from rating, he/she might like:')\n",
    "print('%s %25s %43s' % ('movie_id', 'movie_title', 'movie_genres'))\n",
    "for m, r in prediction[:5]:\n",
    "    mid = movies[m]-1\n",
    "    r = r * 2.5 + 2.5\n",
    "    print('%8s %25s %43s' % (mid, movie_data[mid]['title'][:24], movie_data[mid]['genres'][:4]))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.7.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}